No linear stage, MIND at the image's resolution (no voxel_size), grid 6, search 4, 80 Adam steps, weight 1.25, grid 2, three smoothing passes; itk-impact 0.1.6 and impact-reg-konfai 1.8.7 pinned. A native port of the authors' method, not their code: within 0.03 Dice of their package on the benchmark's check case.
The baselines run through impact-reg-konfai register, like the presets
ANTs_SyN, NiftyReg, Deeds, Greedy, Demons, CorrField, ConvexAdam_MIND, UniGradICON, MultiGradICON, Anatomix_ConvexAdam, SynthMorph, BrainMorph: each folder carries its model.py, which runs the authors' code at a pinned version fetched on first use (a git checkout, a release archive checked by hash, a cmake or make build, or an environment uv makes on its own Python), with the benchmark's settings in Prediction.yml. baselines.py is what they share, copied into each folder by build.py. Each app was checked against the benchmark's own line on one case.
Puts main back as it was at 8acfab54, which impact-reg-konfai 1.8.6 reads. The 1.8.7 files stay at 7fc77d1c, the revision 1.8.7 pins, and come back to main with its release.
Presets for impact-reg-konfai 1.8.7: one IMPACT loss, sized from their declared cost
Sixteen presets narrowed to nine that a user can choose between: the two elastix
intensity ones, the two configurations the IMPACT study recommends (TS/M730 early
layers in Jacobian for CT and CBCT, its last layer with MIND in Static for MR and
CT), the coarse ConvexAdam on MIND and the same pass followed by its refinement,
FireANTs on its own, FireANTs driven by IMPACT, and the anatomix pipeline.
FireANTs_Anatomix reproduces the configuration anatomix's authors register with:
anatomix and MIND features extracted once and registered as feature volumes, each
reduced to its principal components so both fit on one card, on their 6x4x2x1
schedule. Measured on ten PSMAReg pairs: 0.743 Dice against 0.730 for their own
pipeline and 0.145 unregistered. It needs konfai 1.8.6, for the Static mode and
for the rounding anatomix's input requires.
The presets that were dropped were variants of these with another feature model.
Each remaining one writes under its own name, which two of them did not…
24 days ago
feat: no VRAM plan, every preset leaves its patch free and blends its tiles
No app declares a `vram_plan` any more. Every preset leaves its patch free (`patch_size: [0, 0, 0]`): the volume runs as one patch when it fits, and the engine is handed exactly what it was without a patch. Otherwise konfai tiles it on the measured VRAM, the fewest equal patches that fit, overlapping by 20 % and blended with a `Cosinus` window: without a declared overlap and window the last patch written won, which left a seam in the displacement field where a patch starts. A registration forwards one pair: its batch stays 1.
